Use of satellite data in ALADIN/HARMONIE-Norway
Andrea Storto and Roger Randriamampianina
Norwegian Meteorological Institute
PO Box 43 Blindern, N-0313 Oslo, Norway
www.met.no
andrea.storto@met.no, rogerr@met.no ,
Introduction
Understanding the relative impact of the observations is of primary importance for data assimilation community, in order to assess their optimal use in operational systems. In this poster
we present the use of observations in the ALADIN/HARMONIE-Norway data assimilation and forecasting system, their impact by the use of randomization techniques and the impact of
experimental observation types, not yet in the reference system; results from the use of background error covariances from downscaled ensemble analysis are also reviewed. A number
of observations, conventional and satellite, have been assimilated over a large period. Their use is shown in the following table.
Table 1. Use of Observations in the ALADIN-HARMONIE/Norway
Type
Parameter (Channel)
Bias correction
Thinning
TEMP
U, V, T, Q, Z
Only T using ECMWF tables
SYNOP
Z
No
Temporal and spatial
PILOT (Europrof.)
U, V, Z
No
Redundancy check against TEMP
No
Temporal and spatial
Experimental observations
Superobservations derived from CloudSat data:
In this section we present recent results for a number of new
observations type whose assimilation is under evaluation. A strategy for the assimilation of CloudSat CPR observations have been assessed. It consists of a Bayesian analysis
able to produce super-observations of relative humidity, exploiting the ability of CloudSat to detect single and multi layer clouds at very high resolution, both on the vertical and
along the satellite track. NWP data to define the “a priori” knowledge of humidity fields come from ECMWF short-range forecasts. A Monte Carlo approach is then used to
calculate the errors of the super-observations, simulating the Bayesian analysis' errors. 3DVAR correctly assimilates these observations, and a positive impact was found,
especially on mass fields. However, the verification period was rather short, hence, the impact of CloudSat observations over a longer period is under evaluation.
No
DRIBU
Z
AIREP
U, V, T
No
AMV
U, V
No – Use of quality flags
25 Km horizontal
AMSU-A
5 to 13
Air-mass and scan bias correction
80 Km horizontal
AMSU-B, MHS
3, 4, 5
Air-mass and scan bias correction
80 Km horizontal
25 Km horizontal
The system takes also advantage of surface analysis by the use of RH and T measurements at 2 meters and wind measurements at 10 m from synoptic stations. The domain (showed
above) has a resolution of 11 Km for a 405x270 computational grid centred over Norway.
Assessment of a strategy for assimilating IASI radiances is not shown here but presented in a separate poster (by R. Randriamampianina).
Degrees of Freedom for Signal
Degrees of freedom for signal (DFS) indicate the self-sensitivity of analysis to different observation types; they are given by the derivative of the analysis, in observation space, with
respect to the observations, and are sensitive to the weight of the observations and to the observation operator formulation. DFS have been computed perturbing all the observations for
5 independent assimilation cycles. Results show that the most important observations in terms of information content carried into the analysis are the wind observations (AIREP, TEMP
and PILOT, although the latter are only a few). AMSU-A radiances are very important as well, and also AMSU-B show to have a big information content. Humidity observations (from
TEMP) have a great impact on the analysis but there are too little of them in the assimilation system. Addition of SEVIRI data in the system slightly reduced the contribution from all the
other observations, especially from the “dominant” one, i.e. AIREP and AMSU-A. The use of variational bias correction technique to correct the bias of all radiances emphasises the
information content of all the observations since they are better assimilated. This is true not only for satellite data but also for conventional data.
Figure 3. Innovations and residuals statistics for CloudSat observations (mean on the left, st.dev. on the
right) calculated by day (top) and by vertical level (bottom). The plots show that the residuals are unbiased
and that the weight given to the CloudSat observations is large.
Figure 4. Radiosonde verification scores against experiment
with all observations in Table 1 but without CloudSat
assimilation.
SEVIRI data:
First assimilation trials have also been run with SEVIRI Infrared radiances (channels 5, 6 in clear-sky or above mid-level clouds conditions, channels
7, 9 and 10 in clear-sky above sea). An air-mass bias correction scheme was used, not varying latitudinally. provides reasonable analysis increments for WV channels (5 and 6)
but
not
for
the
window
channels.
Impact
is
promising,
especially
at
low
and
middle-level
atmosphere.
Note, that SEV04 refer to reference runs, while SEV02 to the run with SEVIRI data.
Figure 1. Absolute and relative degrees of freedom of different observation
types
Figure 5. Impact of SEVIRI data against observations
Bias
RMSE
RMSE
Sensitivity of forecasts to observations
The impact of observation types on forecasts have been studied perturbing each observation group, rerunning the assimilation and comparing the RMSE (between forecasts and
analyses valid at forecast time using the reference system) from the reference experiment (all the observation in the above table) with the RMSE from the perturbed experiments. The
bigger is the relative variation in RMSE, the more sensitive are the forecasts to the observations group. The perturbation and forecast has been repeated for 4 assimilation cycle, far
enough in time from each other to ensure ergodicity of results. This test was done without observations under investigation.
AMSU-A have showed the biggest impact on the forecasts for almost all the parameters, followed by TEMP, whose impact was very strong in the high atmosphere. AIREP observations
seem very important for short-range forecasts, especially for temperature fields, while AMSU-B exercises have influence mostly on low and high level humidity.
Figure 6. Impact of SEVIRI data against analyses
GPS ZTD data:
A feasibility study for the assimilation of GPS Zenith Total Delay has been carried out trough the creation of a dataset of reliable GPS stations
and data processing centre. This approach leaded to 52 assimilable stations over the Norwegian domain. Trials experiment, that used an overestimation of the delay errors
trough the observations minus guess standard deviations have been performed over one month (Feb. 2008).
A flat bias correction is used. Impact on verification against TEMP is small for all the parameters except for temperature, where it is positive in general.
A more rigorous formulation of observations errors is under evaluation.
Figure 7. Impact of GPS ZTD data on forecast of geopotential and temperature
Figure 2. Absolute percentual variation of RMSE of different observation types
Conclusions
The use of variational bias correction technique to correct the bias of all radiances emphasises the information content of all the
observations since they are better assimilated. This is true not only for satellite data but also for conventional data.
The evaluation of the impact of different observations on analysis and forecasts showed that the sensitivity of the analysis was
particularly high with regards to wind measurements, especially from aircrafts, while humidity observations had a large information
content but were not very dense if compared with other observed parameters.
B covariances from downscaled ensemble analysis
In order to exalt background covariances typical of small scales, ensemble analysis generated at ECMWF (kindly provided by Lars Isaksen)
by perturbing observations in the assimilation, have been used after downscaling as initial and lateral boundary conditions for
ALADIN/HARMONIE-Norway 6 hours forecasts. The experiment used all the original 10 members, for a period of 1 month (two daily runs),
and differences between the true state of the atmosphere and the model forecasts have been simulated as differences between the ensemble
mean and the forecasts coming from each of 10 members initial conditions. Background error covariances follow Loik Berre's formulation.
Results have been extensively compared with background error covariances obtained through the “NMC” method, for a winter three-months
dataset of differences between 48 and 24 hours forecasts. The use of ensemble analysis produced much shorter correlations, especially on
the vertical correlations and at large horizontal scales. Variances and cross-covariances present a very similar structure in the two B statistics.
Analysis initialized by 3DVAR using ensemble-derived statistic result closer to the background, and the forecasts show a better verification
scores after day 1. The two cross-sections on the left show different analysis increments for a 2 K single-obs innovation (brightness
temperature for the channel 9 from NOAA-18), NMC B (left) against Ensemble B (right). Differences in analysis increments are very big.
Verification (right side) compare an NMC-based experiment against an Ensemble-B experiment.
According to our results, AMSU-A data have the most remarkable impact on forecasts, at all the forecast ranges.
A positive impact of CloudSat data – especially on the mass fields - was found against observations in a short period study.
Impact of SEVIRI data on humidity forecasts was found to be positive.
We concluded, that the use of background errors derived through ensemble techniques could contribute to the optimal assimilation of
observations.
Assimilation of other remote-sensed observations types is very promising, although more work for the assessment of their best use
has still to be completed.
Acknowledgements
The authors would like to thank the Meteo-France/GMAP staff for their help, in particular Bernard Chapnik for his suggestions in the computations of DFS and Loik Berre for his
precious comments on the background error statistics. Help from Alena Trojakova (Czech Met Inst) and Gergely Bölöni (Hungarian Met Service) in the assimilation of Seviri data is
also acknowledged. The authors want also to thank Lars Isaksen of ECMWF for having kindly provided the ensemble analysis used for computing background error covariances.
Figure 8. Analysis increments for a 2 K single-obs innovation
(brightness temperature for the channel 9 from NOAA-18)
Figure 9. Comparison of an NMC-based experiment
against an Ensemble-B experiment
Use of satellite data in ALADIN/HARMONIE-Norway
Andrea Storto and Roger Randriamampianina
Norwegian Meteorological Institute
PO Box 43 Blindern, N-0313 Oslo, Norway
www.met.no
andrea.storto@met.no, rogerr@met.no ,
Introduction
Understanding the relative impact of the observations is of primary importance for data assimilation community, in order to assess their optimal use in operational systems. In this poster
we present the use of observations in the ALADIN/HARMONIE-Norway data assimilation and forecasting system, their impact by the use of randomization techniques and the impact of
experimental observation types, not yet in the reference system; results from the use of background error covariances from downscaled ensemble analysis are also reviewed. A number
of observations, conventional and satellite, have been assimilated over a large period. Their use is shown in the following table.
Table 1. Use of Observations in the ALADIN-HARMONIE/Norway
Type
Parameter (Channel)
Bias correction
Thinning
TEMP
U, V, T, Q, Z
SYNOP
Z
PILOT (Europrof.)
U, V, Z
Introduction
Only T using ECMWF tables
Experimental observations
Superobservations derived from CloudSat data:
In this section we present recent results for a number of new
observations type whose assimilation is under evaluation. A strategy for the assimilation of CloudSat CPR observations have been assessed. It consists of a Bayesian analysis
able to produce super-observations of relative humidity, exploiting the ability of CloudSat to detect single and multi layer clouds at very high resolution, both on the vertical and
along the satellite track. NWP data to define the “a priori” knowledge of humidity fields come from ECMWF short-range forecasts. A Monte Carlo approach is then used to
calculate the errors of the super-observations, simulating the Bayesian analysis' errors. 3DVAR correctly assimilates these observations, and a positive impact was found,
especially on mass fields. However, the verification period was rather short, hence, the impact of CloudSat observations over a longer period is under evaluation.
No
No
Temporal and spatial
No
Redundancy check against TEMP
DRIBU
Z
No
Temporal and spatial
AIREP
U, V, T
No
25 Km horizontal
AMV
U, V
No – Use of quality flags
25 Km horizontal
AMSU-A
5 to 13
Air-mass and scan bias correction
80 Km horizontal
AMSU-B, MHS
3, 4, 5
Air-mass and scan bias correction
80 Km horizontal
Experimental observations I.
The system takes also advantage of surface analysis by the use of RH and T measurements at 2 meters and wind measurements at 10 m from synoptic stations. The domain (showed
above) has a resolution of 11 Km for a 405x270 computational grid centred over Norway.
Assessment of a strategy for assimilating IASI radiances is not shown here but presented in a separate poster (by R. Randriamampianina).
Sensitivity of forecasts to observations
The impact of observation types on forecasts have been studied perturbing each observation group, rerunning the assimilation and comparing the RMSE (between forecasts and
analyses valid at forecast time using the reference system) from the reference experiment (all the observation in the above table) with the RMSE from the perturbed experiments. The
bigger is the relative variation in RMSE, the more sensitive are the forecasts to the observations group. The perturbation and forecast has been repeated for 4 assimilation cycle, far
enough in time from each other to ensure ergodicity of results. This test was done without observations under investigation.
Superobservation derived
from CloudSat data
AMSU-A have showed the biggest impact on the forecasts for almost all the parameters, followed by TEMP, whose impact was very strong in the high atmosphere. AIREP observations
seem very important for short-range forecasts, especially for temperature fields, while AMSU-B exercises have influence mostly on low and high level humidity.
Degrees of Freedom for
Signal
Figure 3. Innovations and residuals statistics for CloudSat observations (mean on the left, st.dev. on the
right) calculated by day (top) and by vertical level (bottom). The plots show that the residuals are unbiased
and that the weight given to the CloudSat observations is large.
Figure 4. Radiosonde verification scores against experiment
with all observations in Table 1 but without CloudSat
assimilation.
Experimental observations
SEVIR I data:
First assim ilation trials have also been run w ith SEVIRI Infrared radiances (channels 5, 6 in clear-sky or above m id-level clouds conditions, channels
7, 9 and 10 in clear-sky above sea). An air-m ass bias correction schem e was used, not varying latitudinally. provides reasonable analysis increm ents for W V channels (5 and 6)
but
not
for
the
window
channels.
Im pact
is
prom ising,
especially
at
low
and
m iddle-level
atm osphere.
Note, that SEV04 refer to reference runs, w h ile SEV02 to the run w ith SEVIRI d ata.
Experimental observations II.
SEVIRI data
Figure 5. Impact of SEVIR I data against observations
Figure 2. Absolute percentual variation of RMSE of different observation types
Degrees of Freedom for Signal
Degrees of freedom for signal (DFS) indicate the self-sensitivity of analysis to different observation types; they are given by the derivative of the analysis, in observation space, with
respect to the observations, and are sensitive to the weight of the observations and to the observation operator formulation. DFS have been computed perturbing all the observations for
5 independent assimilation cycles. Results show that the most important observations in terms of information content carried into the analysis are the wind observations (AIREP, TEMP
and PILOT, although the latter are only a few). AMSU-A radiances are very important as well, and also AMSU-B show to have a big information content. Humidity observations (from
TEMP) have a great impact on the analysis but there are too little of them in the assimilation system. Addition of SEVIRI data in the system slightly reduced the contribution from all the
other observations, especially from the “dominant” one, i.e. AIREP and AMSU-A. The use of variational bias correction technique to correct the bias of all radiances emphasises the
information content of all the observations since they are better assimilated. This is true not only for satellite data but also for conventional data.
Figure 6. Impact of SEVIR I data against analy ses
Sensitivity of forecasts to
observations
Experimental observations – III.
Experimental observations III.
GPS ZTD data:
A feasibility study for the assimilation of GPS Zenith Total Delay has been carried out trough the creation of a dataset of reliable GPS stations
and data processing centre. This approach leaded to 52 assimilable stations over the Norwegian domain. Trials experiment, that used an overestimation of the delay errors
trough the observations minus guess standard deviations have been performed over one month (Feb. 2008).
A flat bias correction is used. Impact on verification against TEMP is small for all the parameters except for temperature, where it is positive in general.
A more rigorous formulation of observations errors is under evaluation.
GPS ZTD data
Figure 7. Impact of GPS ZTD data on forecast of geopotential and temperature
Figure 1. Absolute and relative degrees of freedom of different observation
types
Conclusions
The use of variational bias correction technique to correct the bias of all radiances emphasises the information content of all the
observations since they are better assimilated. This is true not only for satellite data but also for conventional data.
Conclusions
The evaluation of the impact of different observations on analysis and forecasts showed that the sensitivity of the analysis was
particularly high with regards to wind measurements, especially from aircrafts, while humidity observations had a large information
content but were not very dense if compared with other observed parameters.
According to our results, AMSU-A data have the most remarkable impact on forecasts, at all the forecast ranges.
B covariances from downscaled ensemble analysis
In order to exalt background covariances typical of small scales, ensemble analysis generated at ECMWF (kindly provided by Lars Isaksen)
by perturbing observations in the assimilation, have been used after downscaling as initial and lateral boundary conditions for
ALADIN/HARMONIE-Norway 6 hours forecasts. The experiment used all the original 10 members, for a period of 1 month (two daily runs),
and differences between the true state of the atmosphere and the model forecasts have been simulated as differences between the ensemble
mean and the forecasts coming from each of 10 members initial conditions. Background error covariances follow Loik Berre's formulation.
Results have been extensively compared with background error covariances obtained through the “NMC” method, for a winter three-months
dataset of differences between 48 and 24 hours forecasts. The use of ensemble analysis produced much shorter correlations, especially on
the vertical correlations and at large horizontal scales. Variances and cross-covariances present a very similar structure in the two B statistics.
Analysis initialized by 3DVAR using ensemble-derived statistic result closer to the background, and the forecasts show a better verification
scores after day 1. The two cross-sections on the left show different analysis increments for a 2 K single-obs innovation (brightness
temperature for the channel 9 from NOAA-18), NMC B (left) against Ensemble B (right). Differences in analysis increments are very big.
Verification (right side) compare an NMC-based experiment against an Ensemble-B experiment.
Comparison of
covariance matrices
A positive impact of CloudSat data – especially on the mass fields - was found against observations in a short period study.
Impact of SEVIRI data on humidity forecasts was found to be positive.
We concluded, that the use of background errors derived through ensemble techniques could contribute to the optimal assimilation of
observations.
Assimilation of other remote-sensed observations types is very promising, although more work for the assessment of their best use
has still to be completed.
Acknowledgements
Acknowledgements
The authors would like to thank the Meteo-France/GMAP staff for their help, in particular Bernard Chapnik for his suggestions in the computations of DFS and Loik Berre for his
precious comments on the background error statistics. Help from Alena Trojakova (Czech Met Inst) and Gergely Bölöni (Hungarian Met Service) in the assimilation of Seviri data is
also acknowledged. The authors want also to thank Lars Isaksen of ECMWF for having kindly provided the ensemble analysis used for computing background error covariances.
Figure 8. Analysis increments for a 2 K single-obs innovation
(brightness temperature for the channel 9 from NOAA-18)
Figure 9. Comparison of an NMC-based experiment
against an Ensemble-B experiment
Use of satellite data in ALADIN/HARMONIE-Norway
Andrea Storto and Roger Randriamampianina
Norwegian Meteorological Institute
PO Box 43 Blindern, N-0313 Oslo, Norway
www.met.no
andrea.storto@met.no, rogerr@met.no ,
Introduction
Understanding the relative impact of the observations is of primary importance for data assimilation community, in order to assess their optimal use in operational systems. In this poster
we present the use of observations in the ALADIN/HARMONIE-Norway data assimilation and forecasting system, their impact by the use of randomization techniques and the impact of
experimental observation types, not yet in the reference system; results from the use of background error covariances from downscaled ensemble analysis are also reviewed. A number
of observations, conventional and satellite, have been assimilated over a large period. Their use is shown in the following table.
Table 1. Use of Observations in the ALADIN-HARMONIE/Norway
Type
Parameter (Channel)
Bias correction
TEMP
U, V, T, Q, Z
Only T using ECMWF tables
No
SYNOP
Z
No
Temporal and spatial
PILOT (Europrof.)
U, V, Z
No
Redundancy check against TEMP
No
Temporal and spatial
Experimental observations
Superobservations derived from CloudSat data:
In this section we present recent results for a number of new
observations type whose assimilation is under evaluation. A strategy for the assimilation of CloudSat CPR observations have been assessed. It consists of a Bayesian analysis
able to produce super-observations of relative humidity, exploiting the ability of CloudSat to detect single and multi layer clouds at very high resolution, both on the vertical and
along the satellite track. NWP data to define the “a priori” knowledge of humidity fields come from ECMWF short-range forecasts. A Monte Carlo approach is then used to
calculate the errors of the super-observations, simulating the Bayesian analysis' errors. 3DVAR correctly assimilates these observations, and a positive impact was found,
especially on mass fields. However, the verification period was rather short, hence, the impact of CloudSat observations over a longer period is under evaluation.
Thinning
DRIBU
Z
AIREP
U, V, T
No
AMV
U, V
No – Use of quality flags
25 Km horizontal
AMSU-A
5 to 13
Air-mass and scan bias correction
80 Km horizontal
25 Km horizontal
The system takes also advantage of surface analysis by the use of RH and T measurements at 2 meters and wind measurements at 10 m from synoptic stations. The domain (showed
AMSU-B,
MHS of 11 Km for
3, a4,405x270
5
Air-mass and
scan
bias correction
80 Km horizontal
above) has
a resolution
computational
grid
centred
over Norway.
Assessment of a strategy for assimilating IASI radiances is not shown here but presented in a separate poster (by R. Randriamampianina).
Degrees of Freedom for Signal
Degrees of freedom for signal (DFS) indicate the self-sensitivity of analysis to different observation types; they are given by the derivative of the analysis, in observation space, with
respect to the observations, and are sensitive to the weight of the observations and to the observation operator formulation. DFS have been computed perturbing all the observations for
5 independent assimilation cycles. Results show that the most important observations in terms of information content carried into the analysis are the wind observations (AIREP, TEMP
and PILOT, although the latter are only a few). AMSU-A radiances are very important as well, and also AMSU-B show to have a big information content. Humidity observations (from
TEMP) have a great impact on the analysis but there are too little of them in the assimilation system. Addition of SEVIRI data in the system slightly reduced the contribution from all the
other observations, especially from the “dominant” one, i.e. AIREP and AMSU-A. The use of variational bias correction technique to correct the bias of all radiances emphasises the
information content of all the observations since they are better assimilated. This is true not only for satellite data but also for conventional data.
Figure 3. Innovations and residuals statistics for CloudSat observations (mean on the left, st.dev. on the
right) calculated by day (top) and by vertical level (bottom). The plots show that the residuals are unbiased
and that the weight given to the CloudSat observations is large.
Figure 4. Radiosonde verification scores against experiment
with all observations in Table 1 but without CloudSat
assimilation.
SEVIRI data:
First assimilation trials have also been run with SEVIRI Infrared radiances (channels 5, 6 in clear-sky or above mid-level clouds conditions, channels
7, 9 and 10 in clear-sky above sea). An air-mass bias correction scheme was used, not varying latitudinally. provides reasonable analysis increments for WV channels (5 and 6)
but
not
for
the
window
channels.
Impact
is
promising,
especially
at
low
and
middle-level
atmosphere.
Note, that SEV04 refer to reference runs, while SEV02 to the run with SEVIRI data.
Figure 5. Impact of SEVIRI data against observations
Bias
RMSE
RMSE
Figure 1. Absolute and relative degrees of freedom of different observation
types
Sensitivity of forecasts to observations
The impact of observation types on forecasts have been studied perturbing each observation group, rerunning the assimilation and comparing the RMSE (between forecasts and
analyses valid at forecast time using the reference system) from the reference experiment (all the observation in the above table) with the RMSE from the perturbed experiments. The
bigger is the relative variation in RMSE, the more sensitive are the forecasts to the observations group. The perturbation and forecast has been repeated for 4 assimilation cycle, far
enough in time from each other to ensure ergodicity of results. This test was done without observations under investigation.
AMSU-A have showed the biggest impact on the forecasts for almost all the parameters, followed by TEMP, whose impact was very strong in the high atmosphere. AIREP observations
seem very important for short-range forecasts, especially for temperature fields, while AMSU-B exercises have influence mostly on low and high level humidity.
Figure 6. Impact of SEVIRI data against analyses
GPS ZTD data:
A feasibility study for the assimilation of GPS Zenith Total Delay has been carried out trough the creation of a dataset of reliable GPS stations
and data processing centre. This approach leaded to 52 assimilable stations over the Norwegian domain. Trials experiment, that used an overestimation of the delay errors
trough the observations minus guess standard deviations have been performed over one month (Feb. 2008).
A flat bias correction is used. Impact on verification against TEMP is small for all the parameters except for temperature, where it is positive in general.
A more rigorous formulation of observations errors is under evaluation.
Figure 7. Impact of GPS ZTD data on forecast of geopotential and temperature
Figure 2. Absolute percentual variation of RMSE of different observation types
Conclusions
The use of variational bias correction technique to correct the bias of all radiances emphasises the information content of all the
observations since they are better assimilated. This is true not only for satellite data but also for conventional data.
The evaluation of the impact of different observations on analysis and forecasts showed that the sensitivity of the analysis was
particularly high with regards to wind measurements, especially from aircrafts, while humidity observations had a large information
content but were not very dense if compared with other observed parameters.
B covariances from downscaled ensemble analysis
In order to exalt background covariances typical of small scales, ensemble analysis generated at ECMWF (kindly provided by Lars Isaksen)
by perturbing observations in the assimilation, have been used after downscaling as initial and lateral boundary conditions for
ALADIN/HARMONIE-Norway 6 hours forecasts. The experiment used all the original 10 members, for a period of 1 month (two daily runs),
and differences between the true state of the atmosphere and the model forecasts have been simulated as differences between the ensemble
mean and the forecasts coming from each of 10 members initial conditions. Background error covariances follow Loik Berre's formulation.
Results have been extensively compared with background error covariances obtained through the “NMC” method, for a winter three-months
dataset of differences between 48 and 24 hours forecasts. The use of ensemble analysis produced much shorter correlations, especially on
the vertical correlations and at large horizontal scales. Variances and cross-covariances present a very similar structure in the two B statistics.
Analysis initialized by 3DVAR using ensemble-derived statistic result closer to the background, and the forecasts show a better verification
scores after day 1. The two cross-sections on the left show different analysis increments for a 2 K single-obs innovation (brightness
temperature for the channel 9 from NOAA-18), NMC B (left) against Ensemble B (right). Differences in analysis increments are very big.
Verification (right side) compare an NMC-based experiment against an Ensemble-B experiment.
According to our results, AMSU-A data have the most remarkable impact on forecasts, at all the forecast ranges.
A positive impact of CloudSat data – especially on the mass fields - was found against observations in a short period study.
Impact of SEVIRI data on humidity forecasts was found to be positive.
We concluded, that the use of background errors derived through ensemble techniques could contribute to the optimal assimilation of
observations.
Assimilation of other remote-sensed observations types is very promising, although more work for the assessment of their best use
has still to be completed.
Acknowledgements
The authors would like to thank the Meteo-France/GMAP staff for their help, in particular Bernard Chapnik for his suggestions in the computations of DFS and Loik Berre for his
precious comments on the background error statistics. Help from Alena Trojakova (Czech Met Inst) and Gergely Bölöni (Hungarian Met Service) in the assimilation of Seviri data is
also acknowledged. The authors want also to thank Lars Isaksen of ECMWF for having kindly provided the ensemble analysis used for computing background error covariances.
Figure 8. Analysis increments for a 2 K single-obs innovation
(brightness temperature for the channel 9 from NOAA-18)
Figure 9. Comparison of an NMC-based experiment
against an Ensemble-B experiment